Mechanism-data fusion driven substation digital twin modeling method and device

Abstract

The application provides a mechanism-data fusion driven digital twin modeling method and device of a transformer substation, and relates to the technical field of transformer substations. The digital twin model of the transformer substation comprises an in-station model and an out-station model. For the out-station system, an out-station steady-state equivalent model is established based on steady-state measurement data, representing the steady-state characteristics of the external power grid. An out-station dynamic response model is established based on dynamic measurement data using a deep learning model, representing the disturbance characteristics of the external power grid. The out-station model can be applied to nonlinear and mechanism-unknown prediction scenarios, improving the accuracy of the digital twin model of the transformer substation.

Description

Technical Field

[0001] This invention relates to the field of substation technology, and in particular to a mechanism-data fusion-driven digital twin modeling method and apparatus for substations. Background Technology

[0002] With the deepening of energy transition, building a new power system based on new energy sources has become an important direction for power development. Against this backdrop, the power system is accelerating its transformation and upgrading towards digitalization, informatization, and intelligence, driving the evolution of the traditional power grid into a modern smart grid. As a key hub connecting generation, transmission, and distribution in the power grid, the level of digitalization and intelligence of substations directly affects the operational efficiency and safety of the entire power system.

[0003] Digital twin technology, as a crucial means to support the intelligent development of substations, is driving a shift from traditional operation and maintenance models to a new paradigm of data-driven, intelligent decision-making. By constructing a digital twin model of a substation that is highly consistent with the physical entity, digital twin technology enables real-time perception, dynamic prediction, and intelligent decision-making regarding the system's operational status. It has already demonstrated broad application prospects in areas such as condition monitoring, fault diagnosis, operation optimization, and operation and maintenance management.

[0004] Traditional modeling is typically based on physical laws and engineering knowledge, describing the structural characteristics and input-output relationships of substations using mathematical forms such as differential-algebraic equations to form mechanistic models with good interpretability. However, in practical applications, due to factors such as complex system structures, opaque operating mechanisms, and dynamically changing boundary conditions, pure mechanistic models often fail to meet accuracy requirements when facing problems such as strong nonlinearity and unclear mechanisms. Summary of the Invention

[0005] This invention provides a mechanism-data fusion-driven substation digital twin modeling method and apparatus to solve the problem of low accuracy of pure mechanism models in the prior art and improve the accuracy of substation digital twin models.

[0006] This invention provides a mechanism-data fusion-driven digital twin modeling method for substations, comprising the following steps: An internal model of the substation is constructed based on the substation's internal system. An external steady-state equivalent model of the substation's external system is established based on steady-state measurement data of the substation ports under steady-state operation. A deep learning model is trained based on dynamic measurement data of the substation ports under disturbed operation to obtain an external dynamic response model. The external model is obtained by combining the external steady-state equivalent model and the external dynamic response model. The substation digital twin model includes both the internal model and the external model.

[0007] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the substation-based internal system construction of the substation model includes: constructing models of various electrical equipment in the substation and the topology of each electrical equipment model based on the substation internal system to obtain a first model; and constructing various protection component models based on the first model to obtain the substation model.

[0008] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the method for establishing an off-site steady-state equivalent model of the substation's external system based on the steady-state measurement data of the substation ports under steady-state operation includes: establishing a source-side Thevenin equivalent model based on the steady-state measurement data of the substation ports under steady-state operation; and establishing a load-side Norton equivalent model based on the steady-state measurement data of the substation ports under steady-state operation; wherein the off-site steady-state equivalent model includes the source-side Thevenin equivalent model and the load-side Norton equivalent model.

[0009] According to the present invention, a mechanism-data fusion-driven substation digital twin modeling method is provided, wherein the off-site steady-state equivalent model includes a source-side Thevenin equivalent model and a load-side Norton equivalent model; the off-site dynamic response model is obtained by training a deep learning model based on dynamic measurement data of substation ports under disturbance operation conditions, including: training a differential algebraic neural network based on dynamic measurement data of substation ports under disturbance operation conditions to obtain a source-side differential algebraic neural network; and training a differential algebraic neural network based on dynamic measurement data of substation ports under disturbance operation conditions to obtain a load-side differential algebraic neural network; wherein the off-site dynamic response model includes the source-side differential algebraic neural network and the load-side differential algebraic neural network.

[0010] According to the present invention, a mechanism-data fusion-driven substation digital twin modeling method is provided, wherein the dynamic measurement data includes source-side port voltage and source-side port current; the method of training a differential algebraic neural network based on the dynamic measurement data of the substation port under disturbance operation to obtain a source-side differential algebraic neural network includes: determining the true value of the dynamic voltage of the source side at multiple time points based on the source-side port voltage, the source-side port current, and the source-side Thevenin equivalent model; determining the source-side state variable at a first time point through a first state initialization network based on the true value of the dynamic voltage of the source side at a first time point and the source-side port current; determining the source-side state variable at multiple other time points using a first differential neural network based on the source-side state variable at the first time point; and determining the predicted value of the dynamic voltage of the source side using a first algebraic neural network based on the source-side state variable and the source-side port current. The loss is calculated based on the true value and predicted value of the dynamic voltage on the source side at the same time to obtain the first loss; the first state initialization network, the first differential neural network, and the first algebraic neural network are jointly trained according to the first loss to obtain the source-side differential algebraic neural network, wherein the source-side differential algebraic neural network includes the trained first state initialization network, the first differential neural network, and the first algebraic neural network.

[0011] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the method for determining source-side state variables at multiple other times based on the source-side state variables at a first time moment using a first differential neural network includes: calculating the time derivative of the current source-side state variable using the first differential neural network, wherein the initial value of the current source-side state variable is the source-side state variable at the first time moment; calculating the product of a preset unit time step and the current time derivative to obtain the current source-side state increment; determining the source-side state variable at the next time moment based on the current source-side state variable and the current source-side state increment; updating the current source-side state variable and returning to the execution step: calculating the time derivative of the current source-side state variable using the first differential neural network until a preset termination condition is met to obtain the source-side state variables at multiple other times.

[0012] According to the present invention, a mechanism-data fusion driven digital twin modeling method for substations is provided, wherein the dynamic measurement data is simulation data, and the method further includes: acquiring fault waveform data of the substation port under disturbance operation; and using the fault waveform data to correct the source-side differential algebraic neural network.

[0013] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the method of correcting the source-side differential algebraic neural network using the fault recording data includes: determining the true values ​​of the dynamic recording voltage of the source side at multiple time points based on the fault recording data and the source-side Thevenin equivalent model; obtaining the predicted values ​​of the dynamic recording voltage of the source side at multiple time points using the source-side differential algebraic neural network based on the fault filtering data; calculating a second loss based on the true value and predicted value of the dynamic recording voltage of the source side at the same time point; and correcting the source-side differential algebraic neural network based on the second loss.

[0014] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the step of calculating the second loss based on the true value and predicted value of the dynamic waveform voltage at the source side at the same time to obtain the second loss includes: calculating the second loss based on the following formula: ; in, As the second loss, N S The number of fault recording data. This represents the true value of the dynamic waveform voltage recorded on the source side. This is the predicted value of the dynamic waveform voltage recorded on the source side. The parameters of the current source-side differential algebraic neural network are... The parameters of the source-side differential algebraic neural network before correction are λ and ... All are regularization coefficients.

[0015] According to the present invention, a mechanism-data fusion-driven substation digital twin modeling method is provided, wherein the dynamic measurement data includes load-side port voltage and load-side port current; the method involves training a differential algebraic neural network based on the dynamic measurement data of the substation ports under disturbance operation to obtain a load-side differential algebraic neural network, including: determining the true value of the dynamic current of the load side at multiple time points based on the load-side port voltage, the load-side port current, and the load-side Norton equivalent model; determining the load-side state variables at the second time point through a second state initialization network based on the true value of the dynamic current of the load side at the second time point and the load-side port voltage at the second time point; and based on the second time point... The load-side state variables are determined using a second differential neural network at multiple time points. Based on the load-side state variables and the load-side port voltage, a second algebraic neural network is used to determine the predicted dynamic current value of the load side. A third loss is obtained by calculating the loss based on the true value and predicted dynamic current value of the load side at the same time point. The second state initialization network, the second differential neural network, and the second algebraic neural network are jointly trained based on the third loss to obtain a load-side differential algebraic neural network, wherein the load-side differential algebraic neural network includes the trained second state initialization network, the second differential neural network, and the second algebraic neural network.

[0016] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the method for determining load-side state variables at multiple time points based on load-side state variables at a second time point using a second differential neural network includes: calculating the time derivative of the current load-side state variable using a second differential neural network, wherein the initial value of the current load-side state variable is the load-side state variable at the second time point; calculating the product of a preset unit time step and the current time derivative to obtain the current load-side state increment; determining the load-side state variable at the next time point based on the current load-side state variable and the current load-side state increment; updating the current load-side state variable; and returning to the execution step: calculating the time derivative of the current load-side state variable using a second differential neural network until a preset termination condition is met to obtain load-side state variables at multiple other time points.

[0017] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the dynamic measurement data is simulation data, and the method further includes: acquiring fault waveform data of the substation port under disturbance operation; and using the fault waveform data to correct the load-side differential algebraic neural network.

[0018] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the method of correcting the load-side differential algebraic neural network using the fault recording data includes: determining the true values ​​of the dynamic recording current of the load side at multiple time points based on the fault recording data and the load-side Norton equivalent model; obtaining the predicted values ​​of the dynamic recording current of the load side at multiple time points using the load-side differential algebraic neural network based on the fault filtering data; calculating a fourth loss based on the true value and predicted value of the dynamic recording current of the load side at the same time point; and correcting the load-side differential algebraic neural network based on the fourth loss.

[0019] According to the mechanism-data fusion driven substation digital twin modeling method provided by the present invention, the fourth loss is obtained by calculating the loss based on the true value and predicted value of the dynamic waveform current on the load side at the same time, including: calculating the fourth loss based on the following formula: ; in, As the fourth loss, N l The number of fault recording data. For t k The true value of the dynamic waveform voltage recorded on the load side at that time. For t k The predicted value of the dynamic waveform voltage recorded on the load side at that time. The parameters of the current load-side differential algebraic neural network are... The parameters of the load-side differential algebraic neural network before correction are λ and All are regularization coefficients.

[0020] This invention also provides a mechanism-data fusion driven digital twin modeling device for substations, comprising the following modules: The substation model building module is used to build substation models based on the substation's internal systems. The steady-state model construction module is used to establish an off-site steady-state equivalent model of the off-site system of the substation based on the steady-state measurement data of the substation ports under steady-state operation. The dynamic model building module is used to train a deep learning model based on dynamic measurement data of substation ports under disturbance operation to obtain an off-site dynamic response model. The off-site model construction module is used to combine the off-site steady-state equivalent model and the off-site dynamic response model to obtain the off-site model, wherein the substation digital twin model of the substation includes the on-site model and the off-site model.

[0021] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the substation digital twin modeling method driven by the mechanism-data fusion as described above.

[0022] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a substation digital twin modeling method driven by the mechanism-data fusion as described above.

[0023] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a substation digital twin modeling method driven by the mechanism-data fusion described above.

[0024] This invention provides a mechanism-data fusion-driven substation digital twin modeling method and apparatus. It constructs an internal substation model based on the substation's internal system; establishes an external steady-state equivalent model of the substation's external system based on steady-state measurement data of the substation ports under steady-state operation; trains a deep learning model based on dynamic measurement data of the substation ports under disturbed operation to obtain an external dynamic response model; and combines the external steady-state equivalent model and the external dynamic response model to obtain an external model. The substation digital twin model includes both the internal and external models. For the external system, an external steady-state equivalent model is established based on steady-state measurement data to characterize the steady-state characteristics of the external power grid; an external dynamic response model is established using a deep learning model based on dynamic measurement data to characterize the disturbance characteristics of the external power grid. This enables the external model to be applicable to nonlinear prediction scenarios with unclear mechanisms, improving the accuracy of the substation digital twin model. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating a mechanism-data fusion-driven digital twin modeling method for substations provided by the present invention. Figure 2 This is a schematic diagram of the protective element model construction process provided by the present invention; Figure 3 This is a schematic diagram of the present invention providing a method for converting a source-side system into a source-side Thevenin equivalent model; Figure 4 This is a schematic diagram of the present invention providing a method for converting a load-side system into a load-side Norton equivalent model; Figure 5 This is a schematic diagram of the training process of the source-side differential algebraic neural network provided by the present invention; Figure 6 This is a schematic diagram of a digital twin model architecture for substations provided by the present invention; Figure 7 This is an equivalent circuit diagram of a possible substation provided by the present invention; Figure 8 This is a schematic diagram illustrating the structure and uses of the seven models provided by this invention; Figure 9 This is a schematic diagram of the parameter estimation results of the source-side Thevenin equivalent model provided by the present invention; Figure 10 This is a schematic diagram of the parameter estimation results of the load-side Norton equivalent model provided by the present invention; Figure 11 This is a schematic diagram of the root mean square error of the electrical quantity waveforms of various models provided by this invention. Figure 12 This is a schematic diagram of a mechanism-data fusion driven digital twin modeling device for substations provided by the present invention; Figure 13 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] With the advancement of sensing technology and the abundance of data resources, data-driven methods, due to their flexibility and strong expressive power, have gradually become an important supplement to digital twin modeling. Especially in scenarios where the system mechanisms are unknown or difficult to model, data-driven methods demonstrate significant advantages in model completion and accuracy improvement. In recent years, an increasing number of studies have explored data-driven modeling methods based on deep learning. Building upon this, hybrid modeling techniques that integrate physical knowledge and data-driven approaches have gradually become a research hotspot. These methods typically utilize knowledge-driven approaches to represent the structurally well-defined and relationally definite parts of the system, while simultaneously introducing data-driven models to characterize nonlinear and stochastic processes that are difficult to model precisely. This replaces some incomplete mechanistic models, forming a hybrid modeling system with high accuracy, high adaptability, and a certain degree of interpretability, becoming one of the key technical paths for realizing digital twin simulation of power systems.

[0029] For digital twin modeling of substation scenarios, the core objective is to construct a high-fidelity model that can evolve synchronously with the physical system and be continuously updated, in order to support key applications such as fault analysis, operational assessment, and decision support. However, due to limitations in the complexity of the modeling object and data acquisition conditions, existing methods still have certain shortcomings in terms of modeling accuracy and environmental adaptability, mainly reflected in the following aspects: (1) Insufficient coordination between in-station and out-of-station modeling: In-station systems have clear structures and complete information, and can usually achieve fine modeling based on mechanisms; while out-of-station systems, due to their large network scale, complex structure and difficulty in obtaining information, often adopt static equivalent or simplified models, which are difficult to accurately reflect the impact of the external power grid on the dynamic behavior of the substation.

[0030] (2) Data-driven methods are constrained by the scarcity of samples: It is difficult to obtain data on large disturbances and fault conditions, resulting in a limited number of samples for training dynamic equivalent models, which restricts the generalization ability and applicability of the models.

[0031] (3) Insufficient consistency between twin models and actual systems: Due to model simplification and parameter uncertainty, some models are difficult to apply directly to real engineering scenarios, which limits their application effect in actual operation and maintenance and decision-making.

[0032] (4) The model lacks an adaptive update mechanism: When the power grid operation mode or external network structure changes, the existing model is difficult to correct parameters or structure in a timely manner, which can easily lead to the model gradually deviating from the actual system.

[0033] To address at least one of the aforementioned problems, this invention proposes a mechanism-data fusion-driven digital twin modeling method and apparatus for substations. The following describes the method in conjunction with... Figures 1 to 13 Please provide a detailed explanation.

[0034] Figure 1This is one of the flowcharts illustrating the mechanism-data fusion-driven substation digital twin modeling method provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step S101: Construct an internal model based on the substation's internal systems.

[0035] Based on physical laws and engineering knowledge, a corresponding topology model is established according to the various electrical devices in the substation system, thus obtaining the substation model. Specific methods for constructing the substation model can be found in relevant technologies; no specific limitations are made here.

[0036] Step S102: Based on the steady-state measurement data of the substation ports under steady-state operation, establish an external steady-state equivalent model of the substation's external system.

[0037] Steady-state measurement data at substation ports represents port measurement data under steady-state operating conditions (normal operating conditions), and may include at least one of the following: source-side voltage, source-side current, load-side voltage, and load-side current. Here, the source side refers to the power input side of the substation, and the load side refers to the power output side of the substation. Using the steady-state measurement data, an off-site steady-state equivalent model of the substation's external system is established. This off-site steady-state equivalent model represents the relationships between current, voltage, and other measurement data under steady-state operating conditions.

[0038] Step S103: Train the deep learning model based on the dynamic measurement data of the substation port under disturbance operation to obtain the off-site dynamic response model.

[0039] Dynamic measurement data at substation ports represents port measurement data under disturbance operating conditions (abnormal operating conditions), and may include at least one of the following: source-side voltage, source-side current, load-side voltage, and load-side current. A deep learning model is used to learn the correlations between dynamic measurement data, thereby establishing an external dynamic response model for the substation's external system. The external dynamic response model represents the relationships between disturbance current, disturbance voltage, and other data under disturbance operating conditions. Disturbance current refers to the abnormal electrical quantities recorded at substation ports when the substation experiences a sudden external disturbance, with the current and voltage momentarily deviating from their normal values.

[0040] Step S104: Combine the external steady-state equivalent model and the external dynamic response model to obtain the external model, wherein the substation digital twin model of the substation includes the internal model and the external model.

[0041] By combining the off-site steady-state equivalent model and the off-site dynamic response model, the off-site model can be obtained. The off-site steady-state equivalent model represents the steady-state component of the off-site model, while the off-site dynamic response model represents the disturbance component. The resulting digital twin model of the substation includes both the on-site and off-site models, enabling comprehensive digital twin simulation of the substation.

[0042] In this embodiment of the invention, a digital twin model of the substation is established by collaboratively integrating in-station mechanism modeling and out-of-station data-driven modeling. For the out-of-station system, an out-of-station steady-state equivalent model is established based on steady-state measurement data to characterize the steady-state characteristics of the external power grid; an out-of-station dynamic response model is established using a deep learning model based on dynamic measurement data to characterize the disturbance characteristics of the external power grid. This enables the out-of-station model to be applicable to nonlinear prediction scenarios with unclear mechanisms, improving the accuracy of the substation digital twin model.

[0043] For the substation system, two modeling processes can be used: primary system modeling (wiring structure modeling) and secondary system modeling (protection configuration implementation mechanism modeling) to accurately depict the equipment topology and protection control logic. In one possible implementation, in Figure 1 Based on the illustrated embodiment, the substation-based internal system construction model includes: Step A1: Based on the substation's internal system, construct models of each electrical device within the substation and the topology of each electrical device model to obtain the first model.

[0044] Primary system modeling mainly focuses on modeling the main electrical equipment and their topology within a substation to accurately depict electrical characteristics and topological connections. During primary system modeling, based on the substation's primary wiring diagram, design drawings, and equipment parameter tables, appropriate electrical equipment models are selected, and key parameters such as rated voltage, capacity, and impedance characteristics are set. The network topology of each electrical equipment model within the substation is then constructed to obtain the first model. In one example, modeling can be performed using the cloud-based power system simulation platform CloudPSS SimStudio. This platform has a rich set of basic electrical component models, such as transformers, circuit breakers, disconnectors, busbars, cables, current transformers, and voltage transformers, allowing users to configure equipment parameters and connections as needed.

[0045] Step A2: Based on the first model, construct various protection element models to obtain the station model.

[0046] Secondary system modeling primarily focuses on the relay protection, measurement and control, and automation equipment within the substation. By adding various protection element models to the primary model, real-time monitoring and control responses to the primary model are achieved. In one example, CloudPSS SimStudio's module encapsulation function can be used to construct protection element models. These models implement at least one of the following logics: overcurrent protection logic, zero-sequence overcurrent protection logic, differential protection logic, distance protection logic, reclosing logic, and automatic transfer switch logic. Figure 2 As shown, the constructed protection element models consist of two parts: a principle model and an external characteristic model. The principle model, based on relay protection principles, performs logical operations according to measured and set values ​​to complete action judgment and protection decisions. The external characteristic model is designed for signal interaction with the primary system and other equipment, supporting functions such as function configuration, protection setting, and status report output. To achieve complete information interaction, multiple interfaces can be set outside the protection element model to implement functions such as operation command input, equipment status input, measured electrical quantity input, protection action and alarm output, and equipment parameter setting input.

[0047] After completing the basic construction of the protection element model, a Substation Configuration Description (SCD) file can be introduced. The SCD file is an important standardized data source describing the configuration, communication structure, and logical functions of Intelligent Electronic Devices (IEDs), containing complete information such as IED device configuration, logical node definitions, message communication mappings, and control logic relationships. By parsing the SCD file, the protection function configuration, communication link structure, and action triggering logic of each IED can be extracted. Based on this, two types of core messages are constructed: Generic Object Oriented Substation Event (GOOSE) messages and Sampled Value (SV) messages. GOOSE messages are used to transmit control signals such as status variables and trip commands; SV messages are used to transmit high-frequency sampled values ​​such as voltage and current. Combining the functional configuration and setting parameters of the protection element model, a secondary system model (in-station model) is constructed, including action logic, communication interfaces, and information interaction functions, to achieve response linkage between the protection system and the primary system.

[0048] In this embodiment of the invention, the substation system mainly consists of primary system modeling and secondary system modeling. The first model obtained from the primary system modeling handles power transmission and conversion, while the secondary system modeling, by adding protection element models, achieves status monitoring, protection control, and information interaction. By combining detailed design data and operating parameters, the substation system is accurately modeled, enabling linked simulation of the primary and secondary system models, thus realistically reproducing the substation's operating characteristics. This high-fidelity modeling effectively supports equipment operating status simulation, protection action deduction, and fault response analysis, thereby improving the accuracy of the internal model.

[0049] For off-site systems, steady-state measurement data at the ports can be used to identify the Thevenin equivalent parameters on the source side and the Norton equivalent parameters on the load side, thus characterizing the steady-state characteristics of the external power grid. Figure 1 Based on the illustrated embodiment, in one possible implementation, establishing an off-site steady-state equivalent model of the substation's external system based on the substation port steady-state measurement data under steady-state operation includes: Step B1: Based on the steady-state measurement data of the substation ports under steady-state operation, establish a source-side Thevenin equivalent model.

[0050] Thevenin equivalent model is used to simplify complex linear active two-terminal networks, equating them to a series circuit of a voltage source and a resistor, for example... Figure 3 As shown, where V represents the voltage at the source port, |V| is the magnitude of the voltage at the source port, and ∠θ v Let θ be the initial phase of the source-side port voltage, I represent the source-side port current, |I| be the magnitude of the source-side port current, and ∠θ be the initial phase of the source-side port voltage. I This represents the initial phase of the source-side port current. E th The equivalent voltage of the source-side Thevenin equivalent model, |E th | represents the magnitude of the Thevenin equivalent model voltage on the source side, ∠θ E Z represents the initial phase of the source-side Thevenin equivalent model voltage. th =R th +jX th R is the Thevenin equivalent impedance of the source-side Thevenin equivalent model. th X is the equivalent resistance. th Here, j represents the equivalent reactance, P represents active power, and Q represents reactive power. The Thevenin equivalent model is widely used in circuit analysis and design. For details on how to establish the source-side Thevenin equivalent model, please refer to relevant technical documentation.

[0051] Step B2: Based on the steady-state measurement data of the substation ports under steady-state operation, establish a load-side Norton equivalent model. The off-site steady-state equivalent model includes the source-side Thevenin equivalent model and the load-side Norton equivalent model.

[0052] Norton's equivalent model is used to simplify complex linear active two-terminal networks, equating them to a parallel circuit of a current source and a resistor, for example... Figure 4 As shown, where V represents the voltage at the load-side port, |V| is the amplitude of the voltage at the load-side port, and ∠θ v Let I represent the initial phase of the load-side port voltage, I represent the current at the load-side port, |I| represent the amplitude of the load-side port current, and ∠θ I This represents the initial phase of the load-side port current. nor For the equivalent current in the Norton equivalent model on the load side, |I nor | represents the magnitude of the Norton equivalent model current on the load side, ∠θ Inor Y represents the initial phase of the Norton equivalent model current on the load side. nor =G nor +jB nor For the Norton equivalent admittance of the load-side Norton equivalent model, G nor For equivalent conductance, B nor Here, j represents the equivalent susceptance, P represents active power, and Q represents reactive power. The Norton equivalent model is widely used in circuit analysis and design. For details on establishing the load-side Norton equivalent model, please refer to relevant technical documentation.

[0053] In this embodiment of the invention, an equivalent parameter identification method based on the statistical characteristics of the system's random response is employed. By establishing a Thevenin equivalent model on the source side and a Norton equivalent model on the load side, the steady-state characteristics of the off-site system can be accurately described from both the power supply and load sides. This method fully utilizes the measured data of voltage, current, active power, and reactive power at the ports, and dynamically estimates the equivalent parameters through statistical characteristic analysis and sensitivity modeling. Compared to traditional methods that rely on fault disturbances or detection signals, this method requires no additional excitation and can complete high-precision parameter estimation using only environmental data under steady-state conditions. It has good real-time performance and maintains strong robustness and accuracy even under practical conditions such as low signal-to-noise ratio and asynchronous sampling.

[0054] The following example illustrates the process of establishing the source-side Thevenin equivalent model and the load-side Norton equivalent model. In this example, under the source-side Thevenin equivalent circuit model, the port electrical quantities satisfy the following relationship: (1); Where I* represents the conjugate of the I vector. Under the Norton equivalent model on the load side, the port electrical quantities satisfy the following relationship: (2).

[0055] Converting formulas (1) and (2) into amplitude form respectively yields: (3); (4).

[0056] The expressions for voltage amplitude and current amplitude can be derived from formulas (3) and (4) respectively: (5); (6).

[0057] Taking the partial derivatives of the voltage and current amplitudes in formulas (5) and (6) with respect to active and reactive power, respectively, we can obtain the sensitivity parameter expressions for the voltage and current amplitudes with respect to active and reactive power in the source-side Thevenin equivalent model and the load-side Norton equivalent model, respectively: (7); (8).

[0058] Within a small range, changes in voltage and current amplitudes exhibit a linear response to changes in active and reactive power, expressed mathematically as follows: (9).

[0059] To suppress the impact of measurement noise on sensitivity estimation, a sliding window method is employed to extract effective information through local statistical characteristics. Let the measurement time length be T, and the port measurement sampling period be T. S The total number of sampling points is N = T / T S Let the sliding window length be n (i.e., each window contains n consecutive sampling points). To improve data utilization, the sliding step size is set to 1, meaning each window slides one sampling point, resulting in a total of Nn windows. Within each window, for electrical quantities X (including voltage amplitude |V|, current amplitude |I|, active power P, and reactive power Q), the following statistics are calculated: (10); (11); (12).

[0060] in, This represents the mean of the measurement X within the k-th window; This represents the variance of measurement X within the k-th window; This represents the covariance of active power measurement P and reactive power measurement Q within the k-th window.

[0061] Based on the sliding window mean, a sensitivity parameter estimation model is derived using formula (9). Within each window, the following relationships exist: (13).

[0062] Furthermore, to enhance robustness, based on formula (13), the sensitivity relationship based on the sliding window variance is derived: (14).

[0063] A set of overdetermined equations can be constructed from the linear equations extracted from all Nn sliding windows. By performing regression fitting on this set of equations using the least squares method, the optimal estimate of the sensitivity parameters can be obtained. After obtaining the estimated values ​​of the sensitivity parameters on the source side and the load side, nonlinear overdetermined equations for the equivalent parameters are constructed by combining the port voltage, current amplitude, and power measurement data. The source side is solved by simultaneously solving equations (5) and (7), and the load side is solved by simultaneously solving equations (6) and (8). The Levenberg-Marquardt optimization method is used to solve the equations, and the equivalent voltage amplitude and equivalent impedance parameters of the Thevenin equivalent model on the source side, and the equivalent current source amplitude and equivalent admittance parameters of the Norton equivalent model on the load side are obtained. It is understood that the establishment process of the Thevenin equivalent model on the source side and the Norton equivalent model on the load side is only illustrative and is not intended to limit the scope of protection of this invention.

[0064] In this embodiment of the invention, during the steady-state modeling stage, based on the electrical interface characteristics between the substation and the external power grid, steady-state measurement data from the substation ports are used to model the external system using both the source-side Thevenin equivalent and the load-side Norton equivalent models. The source-side Thevenin equivalent model uses equivalent voltage sources and internal resistance parameters to characterize the power supply capacity and voltage support characteristics of the external power grid to the substation, and can better reflect the short-circuit capacity and voltage stiffness of the external grid. The load-side Norton equivalent model uses equivalent current sources and parallel admittance to describe the load aggregation characteristics of the external system, and is more suitable for characterizing the impact of the load side on the current injection behavior of the system. Through statistical analysis of steady-state measurement data such as port voltage and current, equivalent parameters can be estimated online, achieving a simplified and effective characterization of the steady-state characteristics of the external system, and providing a physically consistent benchmark model for subsequent dynamic modeling.

[0065] The Thevenin equivalent model on the source side and the Norton equivalent model on the load side, obtained from steady-state measurement data, can effectively characterize the steady-state characteristics of the external system under normal operating conditions. However, when disturbances such as faults occur inside or outside the substation, the dynamic response behavior of the external system will directly affect the substation port voltage, current, and protection actions. Relying solely on the steady-state equivalent model is insufficient to accurately reflect the dynamic characteristics of the external system during transient processes. Furthermore, the dynamic characteristics of the external system are influenced by multiple factors such as grid scale, operating mode, and topology, exhibiting significant nonlinear and time-varying characteristics, making it difficult to accurately describe using analytical models with fixed structures and constant parameters. Therefore, it is necessary to introduce a model with strong nonlinear expressive capabilities to supplement the modeling of the dynamic response behavior of the external system, based on the steady-state equivalent model, in order to achieve a high-fidelity characterization of the external system's dynamic characteristics.

[0066] In one possible implementation, the off-site dynamic response model can be obtained by training a Neural Differential Algebraic Equations (NDAE). Figure 1 Based on the illustrated embodiment, the process of training a deep learning model using dynamic measurement data of substation ports under disturbance operating conditions to obtain an off-site dynamic response model includes: Step C1: The differential algebraic neural network is trained based on the dynamic measurement data of the substation port under disturbance operation to obtain the source-side differential algebraic neural network.

[0067] When there is enough dynamic measurement data of the substation under disturbance operation, the actual dynamic measurement data of the substation can be used directly for training. However, when the dynamic measurement data of the substation under disturbance operation is insufficient to support training, dynamic measurement data can be obtained through simulation.

[0068] The source-side differential algebraic neural network can be represented as: (15).

[0069] Among them, X S (t) represents the virtual state variable of the source system at time t, used to characterize the dynamic state of the source system that is not explicitly modeled, and is not actually directly measurable; I SP (t) represents the current at the source-side port of the substation at time t; E S (t) represents the equivalent voltage generated by the dynamic voltage source at time t; function N s ( ) represents the parameter θ s A differential neural network is used to describe the dynamic evolution of the state variables of the source-side system; the function Φ s ( ) represents the parameter ξ sAn algebraic neural network is used to calculate the dynamic voltage source output under given system state and port current conditions. This dynamic voltage source is connected in series with a steady-state equivalent voltage source to characterize the dynamic voltage support characteristics of the external system to the substation ports under disturbance conditions.

[0070] Step C2: The differential algebraic neural network is trained based on the dynamic measurement data of the substation port under disturbance operation to obtain the load-side differential algebraic neural network; the external dynamic response model includes the source-side differential algebraic neural network and the load-side differential algebraic neural network.

[0071] The load-side differential algebraic neural network can be expressed as: (16).

[0072] Among them, X l (t) represents the virtual state variable of the load-side system at time t, used to characterize the dynamic state of the load-side system that is not explicitly modeled, and is not directly measurable in practice; V lp (t) represents the port voltage at the load connection node at time t; I l (t) represents the equivalent current injected by the dynamic current source at time t; function N l ( ) represents the parameter θ l Differential neural networks characterize the dynamic evolution of load state variables; function Φ l ( ) represents the parameter ξ l An algebraic neural network is used to describe the current injection characteristics under given system state and port voltage conditions. This dynamic current source is connected in parallel with the steady-state equivalent current source to characterize the dynamic current response behavior of the off-site system during disturbances.

[0073] In this embodiment of the invention, the execution order of steps C1 and C2 is not limited. Step C1 can be executed first and then step C2, or step C2 can be executed first and then step C1, or steps C1 and C2 can be executed simultaneously. All of these are within the protection scope of this invention.

[0074] In this embodiment of the invention, for the off-site system, a Thevenin equivalent model on the source side and a Norton equivalent model on the load side are established using port steady-state measurement data to characterize the steady-state characteristics of the off-site system. Based on these, a differential algebraic neural network is introduced to describe the dynamic characteristics under disturbed operating conditions, characterizing the dynamic response of the off-site system. Thus, the behavior of the off-site system is decomposed into a steady-state equivalent part and a dynamic response part. The steady-state equivalent part ensures the physical consistency of the model under normal operating conditions, while the dynamic response part compensates for the dynamic deviation of the steady-state equivalent part during transient processes, thereby achieving a unified description of the dynamic characteristics of the off-site system. Through the modeling method of this embodiment, while retaining the clear physical meaning of the traditional equivalent model, dynamic modeling with strong nonlinear expressive capabilities is introduced, achieving efficient characterization of the dynamic characteristics of the off-site system. The training process of the source-side differential algebraic neural network is illustrated below. In one possible implementation, see [link to relevant documentation]. Figure 5 , Figure 5 This is a flowchart illustrating the training process of the source-side differential algebraic neural network in this invention, including: Step S501: Determine the true values ​​of the dynamic voltage on the source side at multiple time points based on the source-side port voltage, the source-side port current, and the source-side Thevenin equivalent model.

[0075] Based on the source-side Thevenin equivalent model, the dynamic measurement data is decomposed forward, where the dynamic measurement data includes source-side port voltage and source-side port current. The source-side port voltage can be expressed as the sum of the source-side Thevenin equivalent model and the true value of the dynamic voltage, i.e.: V sp (t)=E th (t)-Z th I sp (t)+E s (t) (17); Among them, V sp (t) and I sp (t) represents the source-side port voltage and source-side port voltage and current at time t, respectively. th (t) and Z th E represents the equivalent voltage and equivalent impedance of the source-side Thevenin equivalent model at time t. s (t) represents the true value of the dynamic voltage at time t.

[0076] Step S502: Based on the true value of the dynamic voltage on the source side and the source side port current at the first moment, determine the source side state variables at the first moment through the first state initialization network.

[0077] Considering that the source-side state variables in the first differential neural network are virtual states and do not have directly measurable physical counterparts, a first-state initialization network is introduced. The source-side state variables at the first moment are determined based on the true value of the dynamic voltage and the source-side port current at the first moment. The expression for the source-side state variables at the first moment can be: (18); Wherein, the function ψ s () represents the parameter μ s The first state initializes the network, E s (0) represents the true value of the dynamic voltage at the first moment, I sp (0) represents the source-side port current at the first moment.

[0078] Step S503: Based on the source-side state variables at the first time step, use the first differential neural network to determine the source-side state variables at multiple other time steps.

[0079] By differentiating the source-side state variables using a first differential neural network to obtain their derivatives, and then performing increment / decrement calculations at unit time steps, the source-side state variables for the next / previous time step can be obtained. Furthermore, based on these source-side state variables, the first differential neural network can be used to obtain the source-side state variables for other time steps.

[0080] In one example, taking incremental calculation as an example, the determination of source-side state variables at multiple other times based on the source-side state variables at the first time step using a first differential neural network includes: Step D1: For the current source-side state variable, calculate its time derivative using the first differential neural network, where the initial value of the current source-side state variable is the source-side state variable at the first time step.

[0081] The time derivative of the current source-side state variable is calculated using the first differential neural network: (19).

[0082] Step D2: Calculate the product of the preset unit time step and the current time derivative to obtain the current source-side state increment.

[0083] Step D3: Determine the source-side state variables for the next time step based on the current source-side state variables and the current source-side state increment.

[0084] Explicit time integration is performed using a differentiable numerical integrator (such as Euler or Runge-Kutta) to obtain the source-side state variables at the next time step: (20); in, t represents the unit time step.

[0085] Step D4: Update the current source-side state variable and return to the execution steps: For the current source-side state variable, use the first differential neural network to calculate its time derivative until the preset termination condition is met, and obtain the source-side state variables at multiple other time points.

[0086] The preset termination condition can be customized according to the actual situation. For example, it can be to reach a preset number of cycles, or to complete the calculation of the source-side state variables at each time corresponding to the dynamic measurement data.

[0087] Step S504: Based on the source-side state variables and the source-side port current, determine the predicted dynamic voltage value of the source side using a first algebraic neural network.

[0088] The dynamic voltage prediction value at the corresponding time moment is calculated using a first-generation algebraic neural network: (twenty one).

[0089] Step S505: Based on the true value and predicted value of the dynamic voltage on the source side at the same time, the loss is calculated to obtain the first loss.

[0090] The loss function in this step can be customized according to the actual situation. In one example, it can be L2 loss: (twenty two).

[0091] Among them, E S (t k ) for t k The true value of the dynamic voltage on the source side at that time. ) for t k The predicted dynamic voltage value at the source side at time N is the number of true dynamic voltage values.

[0092] Step S506: Jointly train the first state initialization network, the first differential neural network, and the first algebraic neural network according to the first loss to obtain the source-side differential algebraic neural network, wherein the source-side differential algebraic neural network includes the trained first state initialization network, the first differential neural network, and the first algebraic neural network.

[0093] The training process implements end-to-end backpropagation under the automatic differentiation framework. The adaptive moment estimation (Adam) adaptive gradient descent algorithm can be used to jointly optimize the parameters of the first state initialization network, the first differential neural network, and the first algebraic neural network, thereby obtaining the source-side differential algebraic neural network.

[0094] The following example illustrates the training process of a load-side differential algebraic neural network. In one possible implementation, the training method includes: Step E1: Determine the true values ​​of the dynamic current on the load side at multiple time points based on the load-side port voltage, the load-side port current, and the load-side Norton equivalent model. Based on the Norton equivalent model of the load side, the dynamic measurement data is decomposed forward, including the load side port voltage and load side port current. The load side port current is decomposed into the sum of the load side Norton equivalent model value and the true value of the dynamic current, i.e.: I lp (t)=I nor (t)+Y nor V lp (t)+I l (t) (23); Where V lp (t) and I lp (t) represent the load-side port voltage and load-side port current at time t, respectively. nor (t) and Y nor These are the equivalent current and equivalent admittance of the load-side steady-state Norton equivalent model at time t, respectively. l (t) represents the true value of the dynamic current at time t.

[0095] Step E2: Based on the true value of the dynamic current on the load side and the load side port voltage at the second moment, determine the load side state variables at the second moment through the second state initialization network.

[0096] Considering that the load-side state variables in the second differential neural network are virtual states and do not have directly measurable physical counterparts, a second-state initialization network is introduced. Based on the true value of the dynamic current on the load side and the load-side port voltage at the second time step, the load-side state variables at the second time step are determined. The expression for the load-side state variables at the second time step can be: (twenty four); Wherein, the function ψ l ( ) represents the parameter. μ l The second state initializes the network, I l (0) represents the true value of the dynamic current at the second moment, V lp (0) is the load-side port voltage at the second moment.

[0097] Step E3: Based on the load-side state variables at the second time point, the second differential neural network is used to determine the load-side state variables at multiple time points.

[0098] By differentiating the load-side state variables using a second differential neural network, the derivatives of the load-side state variables are obtained. Based on this, increment / decrement calculations are performed according to a unit time step to obtain the load-side state variables for the next / previous time step. Furthermore, based on the load-side state variables for the next / previous time step, the load-side state variables for other time steps can be obtained using the second differential neural network.

[0099] In one example, taking incremental calculation as an example, the determination of load-side state variables at multiple time points based on the load-side state variables at the second time point using a second differential neural network includes: Step E31: For the current load-side state variable, calculate its time derivative using the second differential neural network, where the initial value of the current load-side state variable is the load-side state variable at the second time step.

[0100] The time derivative of the current load-side state variables is calculated using the second differential neural network: (25).

[0101] Step E32: Calculate the product of the preset unit time step and the current time derivative to obtain the current load-side state increment.

[0102] Step E33: Determine the load-side state variables for the next time step based on the current load-side state variables and the current load-side state increment.

[0103] Explicit time integration is performed using a differentiable numerical integrator (such as Euler or Runge-Kutta) to obtain the load-side state variables at the next time step: (26); in, t represents the unit time step.

[0104] Step E34: Update the current load-side state variables and return to the execution steps: For the current load-side state variables, use the second differential neural network to calculate their time derivatives until the preset termination condition is met, and obtain the load-side state variables at multiple other times.

[0105] The preset termination condition can be customized according to the actual situation. For example, it can be to reach a preset number of cycles, or to complete the calculation of load-side state variables at each time corresponding to the dynamic measurement data.

[0106] Step E4: Based on the load-side state variables and the load-side port voltage, determine the predicted dynamic current value of the load side using a second algebraic neural network.

[0107] The dynamic voltage prediction value at the corresponding time point is calculated using a second-generation algebraic neural network: (27).

[0108] Step E5: Based on the true value and predicted value of the dynamic current on the load side at the same time, the loss is calculated to obtain the third loss.

[0109] The loss function in this step can be customized according to the actual situation. In one example, it can be L2 loss: (28).

[0110] Among them, I l (t k ) for t k The true value of the dynamic current on the load side at that time. ) for t k The predicted dynamic current value on the load side at a given time, where N is the number of true dynamic current values.

[0111] Step E6: Jointly train the second state initialization network, the second differential neural network, and the second algebraic neural network according to the third loss to obtain the load-side differential algebraic neural network, wherein the load-side differential algebraic neural network includes the trained second state initialization network, the second differential neural network, and the second algebraic neural network.

[0112] The training process implements end-to-end backpropagation under the automatic differentiation framework. The adaptive moment estimation and adaptive gradient descent algorithm can be used to jointly optimize the parameters of the second-state initialization network, the second differential neural network, and the second algebraic neural network, thereby obtaining the load-side differential algebraic neural network.

[0113] In this embodiment of the invention, to accurately describe the dynamic response behavior of the external system under disturbance conditions, a source-side / load-side differential algebraic neural network is introduced to model the dynamic characteristics of the equivalent external model. The source-side / load-side differential algebraic neural network combines neural networks with differential equations, effectively approximating complex and nonlinear dynamic processes while maintaining the continuity of system state evolution. It is suitable for the diverse dynamic characteristics exhibited by the external power grid under different operating modes.

[0114] In real-world scenarios, large disturbance events occur infrequently, resulting in a scarcity of real-world samples. To obtain sufficient dynamic measurement data, in one possible embodiment, an operation mode model is introduced, using a dispatch center for power grid operation mode calculation and safety verification. The dispatch center, built upon the large power grid structure, includes network topology and equipment parameter information, enabling flexible setting of typical fault conditions and transient simulation analysis. It is suitable for generating dynamic measurement data under large disturbance conditions. Based on the operation mode model, multiple transient simulation conditions are constructed for both the source-side and load-side systems, generating dynamic measurement data through batch simulation. During the simulation, various typical fault types are set, including one or more of line N-1, single-phase short circuit, two-phase short circuit, two-phase grounding, and three-phase grounding, to cover different disturbance intensities and fault characteristics. For source-side system fault scenarios, dynamic measurement data such as voltage and current at the load-side ports are recorded for training the load-side differential algebraic neural network; for load-side system network fault scenarios, dynamic response data at the source-side ports are collected for training the source-side differential algebraic neural network.

[0115] When all dynamic measurement data are simulation data, it is also necessary to use real measurement data (hereinafter referred to as fault waveform data to distinguish it from the dynamic measurement data obtained by simulation) to verify the trained source-side differential algebraic neural network and load-side differential algebraic neural network.

[0116] In one possible implementation, for the source-side differential algebraic neural network: acquire fault waveform data of the substation ports under disturbance operation; and use the fault waveform data to correct the source-side differential algebraic neural network. The fault waveform data is data obtained from actual measurements of the substation ports under disturbance operation, and its specific content is the same as that of dynamic measurement data, the only difference being that dynamic measurement data is obtained through simulation, while fault waveform data is obtained through actual measurement.

[0117] The process of correcting the source-side differential algebraic neural network using the fault recording data may include: Step F1: Based on the fault recording data and the source-side Thevenin equivalent model, determine the true value of the dynamic recording voltage on the source side at multiple time points.

[0118] The specific implementation process of step F1 can be found in step S501 above, and will not be repeated here.

[0119] Step F2: Based on the fault filtering data, the predicted dynamic waveform voltage values ​​of the source side at multiple time points are obtained using the source-side differential algebraic neural network.

[0120] The specific implementation process of step F2 can be found in steps S502-S504 above, and will not be repeated here.

[0121] Step F3: Based on the true value and predicted value of the dynamic recording voltage on the source side at the same time, the loss is calculated to obtain the second loss.

[0122] The second loss can be L2 loss. However, in order to make the fault recording data have a better correction effect, the loss function used in step F3 can be optimized.

[0123] In the process of training the source-side differential algebraic neural network using simulated dynamic measurement data, the domain optimization problem is as follows: (29).

[0124] in, N represents the parameters of the source-side differential algebraic neural network before correction. pre The quantity of dynamic measurement data. For t k The true value of the dynamic voltage on the source side at that time. The t output of the source-side differential algebraic neural network before correction k Dynamic voltage prediction value at any given time. These are the parameters of the source-side differential algebraic neural network before correction.

[0125] In the transfer learning phase (TL), i.e., during the correction process, the domain optimization problem is: (30); (31); in, As the second loss, N S The number of fault recording data. For t k The true value of the dynamic waveform voltage recorded on the source side at that time. For t k The predicted value of the dynamic waveform voltage recorded on the source side at that time. The parameters of the current source-side differential algebraic neural network are... The parameters of the source-side differential algebraic neural network before correction are λ and ... All are regularization coefficients.

[0126] Step F4: Correct the source-side differential algebraic neural network according to the second loss.

[0127] Considering the limited number of fault recording data samples, a smaller learning rate can be used in the transfer learning phase (correction process), and an early stopping strategy can be introduced to prevent overfitting, thereby achieving a balance between model stability and fitting accuracy. In one possible implementation, the learning rate used in the joint training of the first state initialization network, the first differential neural network, and the first algebraic neural network based on the first loss is η. spre The learning rate used in the process of correcting the source-side differential algebraic neural network based on the second loss is η. s , where η s =α×η spre , 0 < α « 1. The corrected source-side differential algebraic neural network can more accurately reconstruct the dynamic response process of port voltage and current under typical fault scenarios, effectively improving the applicability and reliability of the model in real operating environments.

[0128] During the correction process, an iterative optimization method based on gradient descent can be used to update the parameters: θ S （i+1） =θ S （i） -η s L S (θ) S (i) ), where θ S （i） L represents the parameters of the source-side differential algebraic neural network after the i-th update. S (θ) S (i) ) is the second loss during training i.

[0129] Fault waveform data can comprehensively record the voltage and current transient response processes of a system under typical disturbance conditions, serving as an important data source reflecting the dynamic behavior of a real system. Compared to simulation data, fault waveform data has a limited scale and uneven scene distribution, but its physical realism is high, effectively compensating for deviations between simulation models and actual systems. Based on this, this embodiment of the invention employs a transfer learning method to correct the off-site dynamic equivalent model that has been pre-trained using simulation data. During the transfer correction process, the source-side Thevenin equivalent model remains unchanged; only the parameters of the constructed source-side differential algebraic neural network are fine-tuned. While maintaining the unchanged structure of the source-side differential algebraic neural network, the parameters are optimized and updated. In this way, the model inherits the general dynamic laws learned from the simulation data while further absorbing the dynamic characteristic information of the real system, thereby improving the accuracy of the source-side differential algebraic neural network and further enhancing the accuracy of the substation digital twin model.

[0130] In one possible implementation, for the load-side differential algebraic neural network: acquire fault recording data of the substation ports under disturbance operation; and use the fault recording data to correct the load-side differential algebraic neural network.

[0131] The process of correcting the load-side differential algebraic neural network using the fault recording data may include: Step G1: Based on the fault recording data and the Norton equivalent model on the load side, determine the true value of the dynamic recording current on the load side at multiple time points.

[0132] The specific implementation process of step G1 can be found in step E1 above, and will not be repeated here.

[0133] Step G2: Based on the fault filtering data, the predicted values ​​of the dynamic waveform current at multiple time points are obtained using the load-side differential algebraic neural network.

[0134] The specific implementation process of step G2 can be found in steps E2-E4 above, and will not be repeated here.

[0135] Step G3: Based on the true value and predicted value of the dynamic waveform current on the load side at the same time, the loss is calculated to obtain the fourth loss.

[0136] The fourth loss can be L2 loss. However, in order to make the fault recording data have a better correction effect, the loss function used in step G3 can be optimized.

[0137] In the process of training the load-side differential algebraic neural network using simulated dynamic measurement data, the domain optimization problem is as follows: (32).

[0138] in, For the parameters of the load-side differential algebraic neural network before correction, N pre The quantity of dynamic measurement data. For t k The true value of the dynamic current on the load side at that time. The output t of the load-side differential algebraic neural network before correction k Predicted dynamic current value at any given time. These are the parameters of the load-side differential algebraic neural network before correction.

[0139] During the calibration process, the domain optimization problem is: (33); (34); in, As the fourth loss, N l The number of fault recording data. For t k The true value of the dynamic waveform current recorded on the load side at that time. For t k The predicted value of the dynamic waveform current recorded on the load side at a given time. The parameters of the current load-side differential algebraic neural network are... The parameters of the load-side differential algebraic neural network before correction are λ and All are regularization coefficients.

[0140] Step G4: Correct the load-side differential algebraic neural network according to the fourth loss.

[0141] Considering the limited number of fault recording data samples, a smaller learning rate can be used in the transfer learning phase (correction process), and an early stopping strategy can be introduced to prevent overfitting, thereby achieving a balance between model stability and fitting accuracy. In one possible implementation, the learning rate used during the joint training of the second state initialization network, the second differential neural network, and the second algebraic neural network based on the third loss is η. lpre The learning rate used in the process of correcting the load-side differential algebraic neural network according to the fourth loss is η. l , where η l =ω×η lpre , 0 < ω«1. The corrected load-side differential algebraic neural network can more accurately reconstruct the dynamic response process of port voltage and current under typical fault scenarios, effectively improving the applicability and reliability of the model in real operating environments.

[0142] During the correction process, an iterative optimization method based on gradient descent can be used to update the parameters: θ l （i+1） =θ l （i） -η l L l (θ) l (i) ), where θ l （i） L represents the parameters of the load-side differential algebraic neural network after the i-th update. l (θ) l (i) ) is the second loss during the fourth loss in training i.

[0143] To address the issues of low frequency of large disturbance events and scarcity of real-world measurement data for training in actual operation, this invention introduces a large power grid operation mode model used by the dispatch center for operation mode calculations. Typical large disturbance faults are set in the model for transient simulation calculations. Batch simulation generates dynamic response data covering multiple fault types and disturbance intensities, providing sufficient large disturbance samples to support the training of the neural differential network. Considering the inevitable modeling errors and differences in operating environments between the simulation model and the actual system, a transfer learning method is further introduced to correct the trained source-side / load-side differential algebraic neural network. Using waveform data collected when faults occur in the actual system, the parameters of the load-side differential algebraic neural network are retrained or fine-tuned, allowing the model to better reflect the dynamic characteristics of the real system while retaining the universal dynamic laws learned from the simulation data. By combining simulation data pre-training with actual data transfer correction, the modeling accuracy and generalization ability of the off-site equivalent model in the real operating environment are effectively improved.

[0144] In this embodiment of the invention, a transfer learning method is employed to correct the off-site dynamic equivalent model that has been pre-trained using simulation data. During the transfer correction process, the load-side Norton equivalent model remains unchanged; only the parameters of the constructed load-side differential algebraic neural network are fine-tuned. While maintaining the structure of the load-side differential algebraic neural network, its parameters are optimized and updated. In this way, the model inherits the general dynamic laws learned from the simulation data while further absorbing the dynamic characteristics of the real system, thereby improving the accuracy of the load-side differential algebraic neural network and further enhancing the accuracy of the substation digital twin model.

[0145] This invention employs a modeling approach that combines steady-state equivalent modeling with dynamic modeling using differential algebraic neural networks, and introduces a model correction mechanism based on actual operating data to improve the applicability of the equivalent model under real-world conditions. For example... Figure 6As shown, in the modeling process, for the substation system, mechanistic modeling is carried out based on the primary wiring structure and secondary protection configuration to accurately depict the equipment topology and protection control logic, establishing the substation's primary and secondary models (i.e., the substation model). For the external system, source-side Thevenin parameters are estimated using port steady-state measurement data, thus establishing a source-side Thevenin equivalent model; load-side Norton parameters are estimated using port steady-state measurement data, thus establishing a load-side Norton equivalent model to characterize the external power grid's steady-state characteristics. A differential algebraic neural network is introduced to model the dynamic response behavior of the external system under disturbance conditions. Simulation samples (corresponding to dynamic measurement data) are used to train the source-side / load-side differential algebraic neural networks; further, combined with actual system fault recording data, transfer learning correction is performed on the source-side / load-side differential algebraic neural networks to obtain the external model. Combining the external and internal models forms a set of external system modeling methods that balance physical consistency and data-driven update capabilities. Among them, E Neural I represents the port voltage predicted by the source-side differential algebraic neural network. Neural This represents the dynamic current prediction value of the load-side differential algebraic neural network.

[0146] The following example, using a 110kV substation in a real-world scenario, illustrates the mechanism-data fusion-driven digital twin modeling method for substations described in this invention. The equivalent circuit diagram of this substation is shown below. Figure 7 As shown, for the substation system, the primary system includes two 110kV incoming lines, two 110kV busbars, three main transformers, two station service transformers, four 10kV busbars, four 10kV outgoing lines, four sets of capacitors, three grounding transformers, and various circuit breakers and disconnectors. The CloudPSS SimStudio graphical modeling environment was used to complete the main wiring and component modeling based on the substation's primary drawings and equipment parameters. The secondary system modeling was based on parsing the SCD file, extracting the communication relationships between IEDs, identifying GOOSE and SV channels, and combining the functional configurations and setting values ​​of each protection device to complete the construction of the protection elements and their operating logic.

[0147] For off-site systems, simulations were performed using the IEEE (Institute of Electrical and Electronics Engineers) standard testing system.

[0148] The source-side system adopts the IEEE 39-node transmission network model to characterize the power supply characteristics and transient dynamic behavior of the upstream power grid. Two 110kV incoming lines from the substation are connected to nodes in the IEEE 39-node system. The load-side system adopts the IEEE 33-node distribution network model with renewable energy access. The four 10kV outgoing lines from the substation are connected to four different IEEE 33-node distribution network models. By adjusting the wiring methods of each distribution network model to create differences in topology and setting up different types and penetration ratios of renewable energy access scenarios, the diversity of load-side network structure and operational characteristics is simulated.

[0149] In this embodiment of the invention, seven model configurations were built on the CloudPSS SimStudio platform. The composition and purpose of each model are as follows: Figure 8 As shown in the figure, the reference model (hereinafter referred to as Model A) serves as a model simulating the "real system" and is used to provide a benchmark for steady-state and transient responses; the operation mode model (hereinafter referred to as Model B) corresponds to the large power grid model used in the dispatch center's operation mode calculation and is used to generate large disturbance simulation samples in batches; the digital twin model (hereinafter referred to as Model C) is the substation digital twin model constructed in this invention; the ideal equivalent model (hereinafter referred to as Model D), the steady-state equivalent model (hereinafter referred to as Model E), and the untransfer-corrected model (hereinafter referred to as Model F) are used to evaluate the differences in steady-state consistency, dynamic response accuracy, and engineering applicability of different off-site modeling strategies, respectively.

[0150] The steady-state measurement data used to establish the source-side Thevenin equivalent model / load-side Norton equivalent model came from the steady-state simulation process of Model A. During the simulation, a small random disturbance was applied to the load side to simulate the natural fluctuation characteristics of the system. The simulation duration was set to 120 seconds, and the sampling frequency was 200 Hz. Time-series data of voltage amplitude V, current amplitude I, active power P, and reactive power Q were collected at the 110 kV incoming port on the source side and each 10 kV outgoing port on the load side. The sliding window length used in calculating the statistical characteristics was set to 5 seconds.

[0151] Figure 9 The parameter estimation results of the Thevenin equivalent model of the source-side system are presented, including the equivalent voltage magnitude and equivalent impedance parameters. The Norton equivalent model of the load side provides parameter estimations for each 10kV outgoing port. Figure 10The equivalent current source amplitude and equivalent admittance parameters corresponding to different outgoing lines are given. The training samples for the differential algebraic neural network are generated by model B. Model B uses a transformer equivalent model to replace the refined mechanism modeling of the primary and secondary systems of the substation, which conforms to the modeling habits of large power grid models in the operation mode calculation and safety verification of the dispatch center, and can efficiently carry out transient simulation of large disturbances. In model B, typical fault traversal is performed on both the source side and the load side system. The source side traverses line N-1 faults and single-phase short circuits, two-phase short circuits, two-phase ground short circuits, and three-phase short circuits at each node, recording the dynamic response of voltage and current at the load side port, generating 463 sets of samples for training and testing the differential algebraic neural network on the load side. The same method is used to generate 1600 sets of fault samples on the load side, and the source side port measurement data is recorded for training and testing the source side dynamic model. All samples are randomly divided into training and validation sets at an 8:2 ratio.

[0152] During training, based on the steady-state Thevenin / Norton equivalent parameters identified above, the dynamic deviation (dynamic voltage prediction / dynamic current prediction) calculated from dynamic measurement data is used as the learning object of the differential algebraic neural network. The mean square error is used as the loss function, and the parameters are trained through the Adam optimizer.

[0153] To simulate a typical fault process in a real-world operating environment, a three-phase short-circuit fault was set up on the 110 kV II line in Model A. After the fault occurred, the differential protection of this line operated correctly, disconnecting the circuit breakers at both ends of the line, causing the 10 kV 3M busbar to lose power support. Subsequently, the automatic transfer switch for the 110 kV busbar backup power supply activated, closing the 110 kV bus tie sectionalizing switch, restoring power supply to the 10 kV 3M line from the 110 kV I line. This process fully reflects the actual operating logic of the protection operation, circuit breaker operation, and power restoration within the substation, while also posing higher modeling requirements for the transient response of the external system.

[0154] During the aforementioned fault and recovery process, voltage and current fault waveform data were collected at the substation ports for transfer learning correction of the differential algebraic neural network model. In the transfer phase, the differential algebraic neural network trained on simulation samples was used as the initial model, maintaining the network structure unchanged and only fine-tuning the network parameters. After transfer learning correction, the updated differential algebraic neural network model was integrated with the steady-state equivalent model to form a complete off-site equivalent model, which was then connected to the substation digital twin simulation system for transient calculations. Specifically, in the source-side system, the source-side differential algebraic neural network was connected in series with the Thevenin equivalent voltage source and equivalent impedance to form a complete source-side equivalent model; in the load-side system, the source-side differential algebraic neural network was combined with the Norton equivalent current source and parallel admittance to form the load-side equivalent model.

[0155] Under the aforementioned fault conditions, the transient response characteristics of models A, C, D, E, and F are compared and analyzed. Model A serves as a reference system, used to characterize the true dynamic behavior of the complete primary and secondary systems and the external network; models D and E represent the idealized model without dynamic modeling and the steady-state equivalent model, respectively; models F and C both incorporate steady-state equivalence and dynamic modeling, with model C further corrected through transfer learning of fault waveform data. To comprehensively evaluate the response capability of each model to the aforementioned system fault process, the following four key electrical quantities are selected for comparative analysis: the effective value of the 110 kV II line voltage, the effective value of the 110 kV II line current, the effective value of the 10 kV 3M bus voltage, and the effective value of the 10 kV 3M outgoing line current. These quantities directly reflect the dynamic changes in the system's electrical state during fault occurrence, protection action, and power restoration. Based on the differences between each model and the reference model, the root mean square error of the waveforms of the above four electrical quantities is calculated, and the results are as follows: Figure 11 As shown.

[0156] It can be concluded that different off-site modeling strategies have significantly different capabilities in characterizing the transient response of the system. Model D uses an ideal voltage source and static load to describe the off-site system, without considering the equivalent impedance and dynamic characteristics of the external power grid. Its voltage and current transient responses deviate significantly from the reference model A, making it difficult to reflect the dynamic behavior of the real system during the fault process. Model E, after introducing Thevenin / Norton steady-state equivalence, can better reflect the steady-state electrical characteristics of the off-site system. The response trend in the initial stage of the fault is basically consistent with the reference model. However, due to the lack of a dynamic modeling mechanism, there are still some errors in the transient oscillation and recovery process after the fault is cleared. In contrast, Model F introduces a differential algebraic neural network for dynamic modeling based on the steady-state equivalent model, enabling the model to simultaneously consider steady-state consistency and dynamic response capability, significantly improving the overall transient response accuracy. Model C further utilizes actual fault recording data for transfer learning correction, making the model's dynamic characteristics closer to the real system. Figure 11 The results show that Model C exhibits the smallest error across the four key electrical quantity indicators selected, and can accurately reproduce the dynamic changes in voltage and current of the reference system during the fault process.

[0157] In summary, waveform comparison and error analysis under typical fault scenarios demonstrate that the mechanism-data fusion-driven substation digital twin modeling method proposed in this invention can more accurately reflect the dynamic operating characteristics of the substation and its external systems, verifying its effectiveness and engineering application potential in complex operating scenarios. In this embodiment, a differential algebraic neural network is introduced to describe the dynamic characteristics under disturbance conditions; furthermore, an event-driven rolling update mechanism is constructed to achieve adaptive updates of the external equivalent model when the power grid operating mode changes. Numerical examples show that this method can effectively improve the modeling accuracy of the external equivalent model for the transient response of substation ports, providing support for substation digital twin modeling and application.

[0158] The mechanism-data fusion-driven substation digital twin modeling device provided by this invention is described below. The mechanism-data fusion-driven substation digital twin modeling device described below can be referred to in correspondence with the mechanism-data fusion-driven substation digital twin modeling method described above. See also... Figure 12 The device includes: The in-station model building module 701 is used to build in-station models based on the in-station system of a substation.

[0159] The steady-state model construction module 702 is used to establish an external steady-state equivalent model of the substation's external system based on the steady-state measurement data of the substation ports under steady-state operation.

[0160] The dynamic model building module 703 is used to train a deep learning model based on dynamic measurement data of substation ports under disturbance operation to obtain an off-site dynamic response model.

[0161] The off-site model construction module 704 is used to combine the off-site steady-state equivalent model and the off-site dynamic response model to obtain the off-site model, wherein the substation digital twin model of the substation includes the on-site model and the off-site model.

[0162] In one possible implementation, the substation model construction module 701 is specifically used for: constructing models of various electrical equipment within the substation and the topology of each electrical equipment model based on the substation's internal system, to obtain a first model; and constructing various protection component models based on the first model to obtain the substation model.

[0163] In one possible implementation, the steady-state model construction module 702 is specifically used to: establish a source-side Thevenin equivalent model based on the steady-state measurement data of the substation ports under steady-state operation; and establish a load-side Norton equivalent model based on the steady-state measurement data of the substation ports under steady-state operation; wherein the off-site steady-state equivalent model includes the source-side Thevenin equivalent model and the load-side Norton equivalent model.

[0164] In one possible implementation, the off-site steady-state equivalent model includes a source-side Thevenin equivalent model and a load-side Norton equivalent model; the station dynamic model construction module 703 includes: The source-side model training submodule is used to train the differential algebraic neural network based on the dynamic measurement data of the substation port under disturbance operation, so as to obtain the source-side differential algebraic neural network.

[0165] The load-side model training submodule is used to train the differential algebraic neural network based on the dynamic measurement data of the substation ports under disturbance operation, so as to obtain the load-side differential algebraic neural network.

[0166] The off-site dynamic response model includes the source-side differential algebraic neural network and the load-side differential algebraic neural network.

[0167] In one possible implementation, the dynamic measurement data includes source-side port voltage and source-side port current; the source-side model training submodule is specifically used for: determining the true dynamic voltage of the source side at multiple time points based on the source-side port voltage, the source-side port current, and the source-side Thevenin equivalent model; determining the source-side state variables at a first time point using a first state initialization network based on the true dynamic voltage of the source side at the first time point and the source-side port current; determining the source-side state variables at multiple other time points using a first differential neural network based on the source-side state variables at the first time point; determining the predicted dynamic voltage of the source side using a first algebraic neural network based on the source-side state variables and the source-side port current; calculating a first loss based on the true dynamic voltage of the source side at the same time point and the predicted dynamic voltage; and jointly training the first state initialization network, the first differential neural network, and the first algebraic neural network based on the first loss to obtain a source-side differential algebraic neural network, wherein the source-side differential algebraic neural network includes the trained first state initialization network, the first differential neural network, and the first algebraic neural network.

[0168] In one possible implementation, the source-side model training submodule is specifically used to calculate the time derivative of the current source-side state variable using a first differential neural network, wherein the initial value of the current source-side state variable is the source-side state variable at the first time step; calculate the product of a preset unit time step and the current time derivative to obtain the current source-side state increment; determine the source-side state variable at the next time step based on the current source-side state variable and the current source-side state increment; update the current source-side state variable, and return to the execution step: calculate the time derivative of the current source-side state variable using the first differential neural network until a preset termination condition is met to obtain source-side state variables at multiple other time steps.

[0169] In one possible implementation, the dynamic measurement data is simulation data, and the source-side model training submodule is further used to: acquire fault waveform data of the substation port under disturbance operation; and use the fault waveform data to correct the source-side differential algebraic neural network.

[0170] In one possible implementation, the source-side model training submodule is specifically used for: determining the true values ​​of the dynamic waveform voltage at multiple time points based on the fault waveform data and the source-side Thevenin equivalent model; obtaining predicted values ​​of the dynamic waveform voltage at multiple time points using the source-side differential algebraic neural network based on the fault filtering data; calculating a second loss based on the true values ​​and predicted values ​​of the dynamic waveform voltage at the same time point; and correcting the source-side differential algebraic neural network based on the second loss.

[0171] In one possible implementation, the source-side model training submodule is specifically used to: calculate the second loss based on the following formula: ; in, As the second loss, N S The number of fault recording data. This represents the true value of the dynamic waveform voltage recorded on the source side. This is the predicted value of the dynamic waveform voltage recorded on the source side. The parameters of the current source-side differential algebraic neural network are... The parameters of the source-side differential algebraic neural network before correction are λ and ... All are regularization coefficients.

[0172] In one possible implementation, the learning rate used during the joint training of the first state initialization network, the first differential neural network, and the first algebraic neural network based on the first loss is η. spreThe learning rate used in the process of correcting the source-side differential algebraic neural network based on the second loss is η. s , where η s =α×η spre , 0 < α«1.

[0173] In one possible implementation, the dynamic measurement data includes load-side port voltage and load-side port current; the load-side model training submodule is specifically used for: determining the true value of the dynamic current of the load side at multiple time points based on the load-side port voltage, the load-side port current, and the load-side Norton equivalent model; determining the load-side state variables at a second time point using a second state initialization network based on the true value of the dynamic current of the load side at a second time point and the load-side port voltage; determining the load-side state variables at multiple time points using a second differential neural network based on the load-side state variables at the second time point; determining the predicted value of the dynamic current of the load side using a second algebraic neural network based on the load-side state variables and the load-side port voltage; calculating a third loss based on the true value and predicted value of the dynamic current of the load side at the same time point; and jointly training the second state initialization network, the second differential neural network, and the second algebraic neural network based on the third loss to obtain a load-side differential algebraic neural network, wherein the load-side differential algebraic neural network includes the trained second state initialization network, the second differential neural network, and the second algebraic neural network.

[0174] In one possible implementation, the load-side model training submodule is specifically used for: calculating the time derivative of the current load-side state variable using a second differential neural network, wherein the initial value of the current load-side state variable is the load-side state variable at the second time step; calculating the product of a preset unit time step and the current time derivative to obtain the current load-side state increment; determining the load-side state variable at the next time step based on the current load-side state variable and the current load-side state increment; updating the current load-side state variable and returning to the execution step: calculating the time derivative of the current load-side state variable using a second differential neural network until a preset termination condition is met, thereby obtaining load-side state variables at multiple other time steps.

[0175] In one possible implementation, the dynamic measurement data is simulation data, and the load-side model training submodule is further used to: acquire fault recording data of the substation port under disturbance operation; and use the fault recording data to correct the load-side differential algebraic neural network.

[0176] In one possible implementation, the load-side model training submodule is specifically used for: determining the true values ​​of the dynamic recording current of the load side at multiple time points based on the fault recording data and the load-side Norton equivalent model; obtaining predicted values ​​of the dynamic recording current of the load side at multiple time points using the load-side differential algebraic neural network based on the fault filtering data; performing loss calculation based on the true values ​​and predicted values ​​of the dynamic recording current of the load side at the same time point to obtain a fourth loss; and correcting the load-side differential algebraic neural network based on the fourth loss.

[0177] In one possible implementation, the load-side model training submodule is specifically used to: calculate the fourth loss based on the following formula: ; in, As the fourth loss, N l The number of fault recording data. For t k The true value of the dynamic waveform voltage recorded on the source side at that time. For t k The predicted value of the dynamic waveform voltage recorded on the source side at that time. The parameters of the current source-side differential algebraic neural network are... The parameters of the source-side differential algebraic neural network before correction are λ and ... All are regularization coefficients.

[0178] In one possible implementation, the learning rate for jointly training the second state initialization network, the second differential neural network, and the second algebraic neural network based on the third loss is η. lpre The learning rate used in the process of correcting the load-side differential algebraic neural network according to the fourth loss is η. l , where η l =ω×η lpre , 0 < ω«1.

[0179] Figure 13 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 13 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute any of the mechanism-data fusion driven substation digital twin modeling methods described in this invention.

[0180] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0181] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute any of the mechanism-data fusion driven substation digital twin modeling methods described in the present invention.

[0182] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the mechanism-data fusion driven substation digital twin modeling method described in any of the present invention.

[0183] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0184] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A mechanism-data fusion-driven digital twin modeling method for substations, characterized in that, include: Constructing an internal system model based on the substation; Based on the steady-state measurement data of the substation ports under steady-state operation, an off-site steady-state equivalent model of the off-site system of the substation is established. A deep learning model is trained based on dynamic measurement data of substation ports under disturbance operation to obtain an off-site dynamic response model. By combining the external steady-state equivalent model and the external dynamic response model, an external model is obtained, wherein the substation digital twin model of the substation includes the internal model and the external model.

2. The method according to claim 1, characterized in that, The off-site steady-state equivalent model includes the source-side Thevenin equivalent model and the load-side Norton equivalent model; The deep learning model is trained based on dynamic measurement data of the substation ports under disturbance operation to obtain an off-site dynamic response model, including: The differential algebraic neural network is trained based on the dynamic measurement data of the substation port under disturbance operation to obtain the source-side differential algebraic neural network. The differential algebraic neural network is trained based on the dynamic measurement data of the substation port under disturbance operation to obtain the load-side differential algebraic neural network. The off-site dynamic response model includes the source-side differential algebraic neural network and the load-side differential algebraic neural network.

3. The method according to claim 2, characterized in that, The dynamic measurement data includes source-side port voltage and source-side port current; The differential algebraic neural network is trained based on dynamic measurement data of the substation ports under disturbed operating conditions to obtain a source-side differential algebraic neural network, including: Based on the source-side port voltage, the source-side port current, and the source-side Thevenin equivalent model, determine the true values ​​of the dynamic voltage at multiple moments on the source side. Based on the true value of the dynamic voltage on the source side and the source side port current at the first moment, the source side state variables at the first moment are determined through the first state initialization network. Based on the source-side state variables at the first time step, the first differential neural network is used to determine the source-side state variables at multiple other time steps. Based on the source-side state variables and the source-side port current, the dynamic voltage prediction value of the source side is determined using a first algebraic neural network. The first loss is obtained by calculating the loss based on the true value and predicted value of the dynamic voltage on the source side at the same time. The first state initialization network, the first differential neural network, and the first algebraic neural network are jointly trained based on the first loss to obtain a source-side differential algebraic neural network, wherein the source-side differential algebraic neural network includes the trained first state initialization network, the first differential neural network, and the first algebraic neural network.

4. The method according to claim 3, characterized in that, The method for determining source-side state variables at multiple other times based on the source-side state variables at the first time step using a first differential neural network includes: For the current source-side state variable, its time derivative is calculated using the first differential neural network, where the initial value of the current source-side state variable is the source-side state variable at the first time step. The current source-side state increment is obtained by multiplying the preset unit time step by the current time derivative. The source-side state variables at the next moment are determined based on the current source-side state variables and the current source-side state increment. Update the current source-side state variable and return to the execution steps: For the current source-side state variable, use the first differential neural network to calculate its time derivative until the preset termination condition is met, and obtain the source-side state variable at multiple other time points.

5. The method according to claim 3, characterized in that, The dynamic measurement data is simulation data, and the method further includes: Obtain fault waveform data of the substation ports under disturbance operation conditions; The source-side differential algebraic neural network is corrected using the fault recording data.

6. The method according to claim 5, characterized in that, The step of using the fault recording data to correct the source-side differential algebraic neural network includes: Based on the fault recording data and the source-side Thevenin equivalent model, the true values ​​of the dynamic recording voltage on the source side at multiple time points are determined. Based on the fault filtering data, the predicted dynamic waveform voltage of the source side at multiple time points is obtained using the source-side differential algebraic neural network. The second loss is obtained by calculating the loss based on the true value and predicted value of the dynamic waveform voltage recorded on the source side at the same time. The source-side differential algebraic neural network is corrected based on the second loss.

7. The method according to claim 6, characterized in that, The loss calculation based on the true value and predicted value of the dynamic recording voltage at the source side at the same time moment yields the second loss, which includes: The second loss is calculated based on the following formula: ; in, As the second loss, N S The number of fault recording data. This represents the true value of the dynamic waveform voltage recorded on the source side. This is the predicted value of the dynamic waveform voltage recorded on the source side. The parameters of the current source-side differential algebraic neural network are... The parameters of the source-side differential algebraic neural network before correction are λ and ... All are regularization coefficients.

8. The method according to claim 2, characterized in that, The dynamic measurement data includes load-side port voltage and load-side port current; The differential algebraic neural network is trained based on dynamic measurement data of substation ports under disturbance operating conditions to obtain a load-side differential algebraic neural network, including: Based on the load-side port voltage, the load-side port current, and the load-side Norton equivalent model, determine the true values ​​of the dynamic current on the load side at multiple time points. Based on the true value of the dynamic current on the load side and the load side port voltage at the second moment, the load side state variables at the second moment are determined through the second state initialization network. Based on the load-side state variables at the second time point, the load-side state variables at multiple time points are determined using the second differential neural network; Based on the load-side state variables and the load-side port voltage, the dynamic current prediction value of the load side is determined using a second algebraic neural network. The third loss is obtained by calculating the loss based on the true value and predicted value of the dynamic current on the load side at the same time. The second state initialization network, the second differential neural network, and the second algebraic neural network are jointly trained according to the third loss to obtain a load-side differential algebraic neural network, wherein the load-side differential algebraic neural network includes the trained second state initialization network, the second differential neural network, and the second algebraic neural network.

9. The method according to claim 8, characterized in that, The load-side state variables based on the second time step are used to determine load-side state variables at multiple time steps using a second differential neural network, including: For the current load-side state variable, its time derivative is calculated using a second differential neural network, where the initial value of the current load-side state variable is the load-side state variable at the second time step. The current load-side state increment is obtained by multiplying the preset unit time step by the current time derivative. The load-side state variables for the next time step are determined based on the current load-side state variables and the current load-side state increment. Update the current load-side state variables and return to the execution steps: For the current load-side state variables, use the second differential neural network to calculate their time derivatives until the preset termination condition is met, and obtain the load-side state variables at multiple other times.

10. A mechanism-data fusion driven digital twin modeling device for substations, characterized in that, include: The substation model building module is used to build substation models based on the substation's internal systems. The steady-state model construction module is used to establish an off-site steady-state equivalent model of the off-site system of the substation based on the steady-state measurement data of the substation ports under steady-state operation. The dynamic model building module is used to train a deep learning model based on dynamic measurement data of substation ports under disturbance operation to obtain an off-site dynamic response model. The off-site model construction module is used to combine the off-site steady-state equivalent model and the off-site dynamic response model to obtain the off-site model, wherein the substation digital twin model of the substation includes the on-site model and the off-site model.

Images